AI Help in the Editor

Five AI helpers live on the page editor screen, and a sixth writes your alt text somewhere else entirely. They behave differently enough that treating them as one thing is the main source of confusion here: one drafts into your document as a tracked suggestion, one is a conversation, and three overwrite a single field and save it on the spot. Knowing which is which tells you what happens to your page and how to undo it.

HelperWhere it livesHow you start itWho can use itWhat it writesHow you undo it
AI Writing AssistantThe page canvas[[Ctrl]]+[[J]], or the sparkle in the block hover pillAnyone who can edit the pageA tracked suggestion in the bodyReject the suggestion card
Assistant tabRight rail, first tabClick the tabAdministrators onlyA conversation, plus any action you approveDepends on the action it took
Generate SummaryPage settings → MetadataThe sparkle beside the Summary labelAnyone who can edit the pageReplaces the Summary field and savesRetype the field
Generate KeywordsPage settings → MetadataThe sparkle beside the Keywords labelAnyone who can edit the pageReplaces the whole keyword set and savesRemove the chips and retype
Generate ImagePage settings → Feature ImageThe sparkle beside the Feature Image labelAnyone who can edit the page, on page types that require a feature imageA generated image you choose to keepPick a different asset
Generate Alt TextAsset library, on an imageThe sparkle beside Alt TextAnyone who can edit assetsReplaces the asset’s alt textRetype the field

Drafting with [[Ctrl]]+[[J]]

[[Ctrl]]+[[J]] — [[Cmd]]+[[J]] on a Mac — opens the AI Writing Assistant dialog. The same dialog opens from the magenta sparkle at the bottom of the hover pill that appears at a block’s right edge, which is worth knowing because that route has a useful side effect: if you have selected nothing, clicking the sparkle selects the whole block first, so the assistant is working on something definite rather than on a bare cursor.

The AI Writing Assistant dialog open over the page canvas with a paragraph selected and highlighted behind it, an instruction typed into the prompt box

The dialog is a prompt box — “What would you like to create or improve?” — with a short list of example prompts under it and a Generate button. [[Ctrl]]+[[Enter]] submits without reaching for the mouse; [[Esc]] closes without sending. Your selection stays visibly highlighted the whole time the dialog is open, so you can see what you are about to change.

What gets sent is your instruction plus three pieces of context: the text you selected, up to five thousand characters before it, and up to five thousand characters after it. That is why “make this shorter” works without you explaining what “this” is, and why the result usually matches the voice of the surrounding page. Also sent, invisibly, is the list of formatting features your page type disallows — see below.

The result does not replace your text. It arrives as a tracked suggestion, attributed to you, exactly as if a colleague had typed it in Suggesting mode: the new text is marked as an addition in the canvas and a card appears in the Comments rail with Accept and Reject.

The page canvas after an AI result lands, showing the inserted text marked as a tracked addition and its accept/reject card in the Comments rail
sequenceDiagram
    autonumber
    actor You
    participant Editor as Page editor
    participant AI as Leed AI
    You->>Editor: Ctrl+J, then an instruction
    Editor->>AI: instruction + selected text + ~5k chars before + ~5k chars after<br/>+ the features this page type disallows
    AI-->>Editor: Leed Markdown
    Editor->>Editor: parse it and insert as a tracked suggestion
    Editor-->>You: an addition in the canvas, a card in the rail
    You->>Editor: Accept, or Reject

That round trip is the answer to the question people ask first — why did my text arrive as a suggestion instead of as text? Because the decision to keep it is yours, and because a suggestion is reversible in one click while an overwrite is not. Suggesting mode and tracked changes explains how to read the card and exactly what Reject restores.

Writing a good instruction

Short, imperative and specific beats polite and vague. The assistant already has the surrounding page; what it does not have is your intent.

Two habits worth forming. Select the text you mean before you press the shortcut, rather than describing where it is. And run one instruction at a time — a suggestion covering a single change is quick to judge and quick to reject, while one covering four changes is neither.

Why it sometimes refuses a format

Ask for a table on a page type that has tables switched off and you will get a list instead, with no error and no apology. That is deliberate: before the assistant writes anything, it is handed the list of formatting features this page type disallows and instructed to produce the closest allowed structure instead — a heading with paragraphs in place of a collapsible block, a blockquote in place of an alert.

So a “refusal” is almost never the model being unhelpful; it is your page type. What each page type lets you format lists the nine switches and where an administrator turns one on.

The Assistant tab

The first tab of the right rail is the Leed assistant, scoped to the page you have open — a running conversation rather than a one-shot instruction, badged Leed AI with the note Scoped to this page above the transcript.

The right rail's Assistant tab mid-conversation on a real page, showing the AI badge on the tab icon, a user turn, an assistant turn and the composerThe Assistant tab as a non-administrator sees it, showing the AI badge and the message that the assistant is not available for this account

What the assistant can actually do across the workspace, and which of its actions stop and wait for your approval, is the subject of how Leed’s AI is guarded.

The sparkle buttons in Page settings

Open the right rail’s Page settings tab and scroll to Metadata. Two of the field labels carry a small magenta sparkle:

  • Generate Summary reads the page body and writes a summary into the field.
  • Generate Keywords reads the page body and proposes a keyword set.

The sparkle shakes while the request is running and the field is disabled until it returns. Both then overwrite what was there and save immediately — Generate Keywords replaces the entire chip set rather than adding to it. Neither asks for confirmation, so if you have a summary you like, copy it somewhere before you press the button. Once the value lands it is an ordinary field: edit it, and your edit saves like any other.

Both sparkles are hidden when you cannot edit the page — on a published page, in Viewing mode, or without write access — because the field is read-only in those states anyway. The fields themselves, and what each one does to your published page, are covered in page settings.

The third sparkle sits on Feature Image, and it appears only on page types that require a feature image — so a documentation author will not see it, and a blog author usually will. It opens a Generate Feature Image dialog seeded with your workspace’s AI theme description (read-only until you flip the Edit switch beside it), an optional palette of brand colors to check, and a prompt box. Generate, then keep or discard the result from the modal that follows. Feature-image generation is the only place image generation is mounted in Leed; AI image generation covers the prompt, the theme and what the generated asset becomes.

Alt text you did not write

Upload an image and Leed writes its alt text for you, automatically, as part of the upload. You do not ask for it and there is no button to press — it is there when the asset appears. Two things follow from that: SVGs are skipped, and if the description is wrong or unhelpful, it is yours to fix.

To rewrite it, open the image in the asset library and use the Generate Alt Text sparkle beside the Alt Text field, or type over it. Uploading images covers the upload path and where the alt text ends up on your site.

What AI in the editor will not do

  • It will not publish, delete, or change permissions or billing. Nothing in this list of helpers touches your live site. Publishing stays a deliberate human act with its own privilege.
  • It will not silently overwrite your body text. Every AI edit to the page body lands as a suggestion for you to accept or reject. The metadata sparkles are the exception, and they are narrow — one field each.
  • It will not invent formatting your page type forbids. It is told the constraints before it writes.
  • It does not know about your unsaved thinking. It sees the document as it stands, not the paragraph you were about to delete.

One thing it does do that surprises people: an edit made by an AI client connected over MCP arrives in your open editor live, as tracked suggestions, with the client shown as another presence in the document. You do not need to reload to see it, and it cannot overwrite what you are typing. Authoring pages over MCP is the other side of that conversation.

Where the rest of the AI surface lives

The editor is one of several places Leed puts AI in front of you. The Leed Assistant page describes what it can reach; and your own AI client can connect to Leed directly rather than through a browser. If you are trying to work out which surface you want, AI in Leed indexes all of them in one place.

ESC